基于改进MambaOut的轻量化稀土矿物快速识别算法

A Lightweight Rapid Recognition Algorithm for Rare Earth Minerals Based on Improved MambaOut

摘要:

针对稀土矿物图像背景干扰强、纹理尺度复杂、矿种区分难度大且现有算法推理速度受限的问题，提出一种轻量化稀土矿物快速识别算法（REMamba）。算法采用双路轻量化特征提取模块与高效多尺度上下文聚合模块的协同结构：前者应用于网络浅层，替换原始深度卷积，削减特征冗余、提升浅层特征提取效率；后者嵌入网络深层，并行融合多尺度上下文信息与全局注意力特征，增强矿物多尺度特征感知。在基于MinDat数据集筛选14类稀土矿物构建专用数据集上开展实验，结果表明：相较于基线算法MambaOut-Femto，本文算法的准确率与F1值分别提升1.52%、1.51%，推理速度提升22.7%，推理延迟仅7.4 ms。消融实验验证了模块协同优化的有效性，该算法兼顾高精度与低延迟，适用于矿区边缘设备的快速识别。

Abstract:

To address the issues of strong background interference, complex texture scales, difficulty in distinguishing mineral types, and limited inference speed in existing algorithms, a lightweight rapid rare earth mineral identification algorithm (REMamba) is proposed. The algorithm adopts a collaborative structure combining dual lightweight feature extraction modules and efficient multi-scale context aggregation modules: The former is applied to shallow layers of networks, replacing raw deep convolution, reducing feature redundancy, and improving the efficiency of shallow feature extraction; The latter is embedded deep into the network, parallelly integrating multi-scale contextual information and global attention features to enhance multi-scale feature perception of minerals. Experiments were conducted on a dedicated dataset for constructing 14 categories of rare earth minerals based on the MinDat dataset. The results show that compared to the baseline algorithm MambaOut-Femto, the accuracy and F1 value of the proposed algorithm improve by 1.52% and 1.51%, respectively, inference speed increases by 22.7%, and inference delay is only 7.4 ms. Ablation experiments verified the effectiveness of module co-optimization, which balances high precision and low latency, making it suitable for rapid identification of edge equipment in mining areas.